toward disentangling the e ect of hydrologic and...

WATER RESOURCES RESEARCH, VOL. ???, XXXX, DOI:10.1029/,

Toward disentangling the effect of hydrologic and1

nitrogen source changes from 1992 to 2001 on2

incremental nitrogen yield in the contiguous United3

States4

Md Jahangir Alam1

and Jonathan L. Goodall2

The goal of this research was to quantify the relative impact of hydrologic and nitrogen source5

changes on incremental nitrogen yield in the contiguous United States. Using nitrogen source6

estimates from various federal data bases, remotely-sensed land use data from the National Land7

Cover Data (NLCD) program, and observed instream loadings from the United States Geologi-8

cal Survey (USGS) National Stream Quality Accounting Network (NASQAN) program, we cali-9

brated and applied the spatially-referenced regression model SPARROW to estimate incremental10

nitrogen yield for the contiguous United States. We ran different model scenarios to separate the11

effects of changes in source contributions from hydrologic changes for the years 1992 and 2001,12

assuming that only state conditions changed and that model coefficients describing the stream13

water-quality response to changes in state conditions remained constant between 1992 and 2001.14

Model results show a decrease of 8.2% in the median incremental nitrogen yield over the period15

of analysis with the vast majority of this decrease due to changes in hydrologic conditions rather16

than decreases in nitrogen sources. For example, when we changed the 1992 version of the model17

to have nitrogen source data from 2001, the model results showed only a small increase in median18

incremental nitrogen yield (0.12%). However, when we changed the 1992 version of the model to19

have hydrologic conditions from 2001, model results showed a decrease of approximately 8.7% in20

median incremental nitrogen yield. We did, however, find notable differences in incremental yield21

D R A F T February 13, 2012, 2:39pm D R A F T

This is an Accepted Manuscript of an article published in Water Resources Research in 2012, available online: doi:10.1029/2011WR010967

X - 2 ALAM AND GOODALL: HYDROLOGIC AND SOURCE CHANGES ON NITROGEN LOADING

estimates for different sources of nitrogen after controlling for hydrologic changes, particularly22

for population related sources. For example, the median incremental yield for population related23

sources increased by 8.4% after controlling for hydrologic changes. This is in contrast to a 2.8%24

decrease in population related sources when hydrologic changes are included in the analysis.25

Likewise we found that median incremental yield from urban watersheds increased by 6.8% after26

controlling for hydrologic changes – in contrast to the median incremental nitrogen yield from27

cropland watersheds, which decreased by 2.1% over the same time period. These results suggest28

that, after accounting for hydrologic changes, population related sources became a more signifi-29

cant contributor of nitrogen yield to streams in the contiguous United States over the period of30

analysis. However, this study was not able to account for the influence of human management31

practices such as improvements in wastewater treatment plants or Best Management Practices32

(BMPs) that likely improved water quality, due to a lack of data for quantifying the impact of33

these practices for the study area.34

Keywords: Water quality, anthropogenic activities, land use, non-linear regression modeling,35

hydrology.36

1. Introduction

Anthropogenic activities are altering the nitrogen-cycle and resulting in increased con-37

tributions of excess nitrogen to aquatic ecosystems [Howarth et al., 1996; Galloway et al.,38

1995; Boyer et al., 2002]. This excess nitrogen can have serious environmental impacts39

including algal blooms that contribute to anoxic and hypoxic conditions in waterbodies40

1Graduate Research Assistant,


ALAM AND GOODALL: HYDROLOGIC AND SOURCE CHANGES ON NITROGEN LOADING X - 3

[NRC , 2000]. The frequency and magnitude of hypoxic areas in coastal waterbodies in the41

United States and abroad have increased during the latter half of the twentieth century42

[Diaz , 2001; Rabalais et al., 2002]. In the Gulf of Mexico in particular, increased nitro-43

gen loading from the Mississippi River has caused eutrophication and chronic seasonal44

hypoxia in the shallow waters of the Louisiana Shelf [Alexander et al., 2000]. Many large45

ecosystems that are now severely stressed by hypoxia face declines in fishery production,46

reductions in species diversity, and changes to food web structures [Diaz , 2001]. Studies47

suggest that future riverine nitrogen export is likely to increase by as much as 24% in re-48

sponse to heavier fertilizer use, expanded corn production to meet the increased demand49

of food and bio-fuel production, and an increase in annual river discharge under future50

climate conditions [Han et al., 2009]. These factors make it important to understand how51

regional-scale anthropogenic activities, in addition to direct changes to nitrogen sources52

and hydrology, will impact the delivery of nitrogen to waterbodies.53

Nitrogen sources and delivery involve interrelationships between human, economic, and54

physical systems. Anthropogenic activities are not only altering the spatial distribution of55

nitrogen sources, but also the hydrologic cycle by changing land use, which in turn changes56

Department of Civil and Environmental

Engineering, University of South Carolina,

Columbia, SC, USA

2Assistant Professor, Department of Civil

and Environmental Engineering, University

of South Carolina, Columbia, SC, USA



evapotranspiration rates, precipitation rates, runoff volumes, infiltration rates, and air57

temperature [Feddema et al., 2005; Foley et al., 2005; Juang et al., 2007]. Application of58

nitrogen, increased in the form of fertilizer, is driven by food, and more recently bio-fuel,59

production needs [Stonestrom et al., 2009]. Human activities are also responsible for the60

change of animal waste input and atmospheric deposition. Nitrogen delivery is dependent61

on a watershed’s ecological and biophysical characteristics including soil properties, stream62

availability, average annual temperature and precipitation [Smith et al., 1997; Jones et al.,63

2001]. For these reasons, quantifying the impacts of anthropogenic nitrogen sources and64

hydrology/climate related changes in the contiguous United States, the main focus of this65

study, remains a major research challange.66

Past data collection efforts in the United States have resulted in spatially detailed in-67

formation on nitrogen sources, the physical environment, and instream nitrogen concen-68

trations. Modelers have leveraged these historical data to understand the relationships69

that drive nitrogen fate and transport at regional spatial scales. The SPAtially Refer-70

enced Regressions On Watershed Attributes (SPARROW) model [Smith et al., 1997] is71

an example of a regional-scale nitrogen fate and transport model that uses a combination72

of physically-based process representations (e.g., mass balances, overland and instream73

losses) and statistical regression to explain observed spatial variability in instream nitro-74

gen concentrations. SPARROW has been applied to the Mississippi River Basin [Alexan-75

der and Smith, 2005; Alexander et al., 2008], the Chesapeake Bay Watershed [Roberts and76

Prince, 2010], the contiguous United States [Smith et al., 1997], and other regions in the77

United States and abroad [Elliott et al., 2005; Hoos and McMahon, 2009]. SPARROW is78

unique as a statistical water quality model because it incorporates spatial referencing of79



the watershed and river network when modeling the transfer of nutrients from the land-80

scape to streams and through the river network. Smith et al. [1997] demonstrated that81

spatial referencing increases model accuracy by reducing commonly encountered problems82

of network sparseness, bias, and basin heterogeneity. Although the model has been widely83

used for regional scale analysis, only one study has used the model to understand tempo-84

ral changes [Alexander et al., 2008], and this study was limited to the Mississippi River85

Basin.86

The objective of this study was to quantify the relative effects of changes in cli-87

mate/hydrology compared to changes in nitrogen source contributions to incremental88

nitrogen yield for the contiguous United States over a decadal time period (1992-2001).89

We applied SPARROW to estimate changes in the delivery of nitrogen to waterbodies90

for the years 1992 and 2001 due to changes in source contributions and hydrological fac-91

tors. We used these two years because of the availability of National Land Cover Dataset92

(NLCD) that includes a land use change product dataset for 1992 and 2001. Using datasets93

described in the following section, we parameterized and then calibrated SPARROW for94

the contiguous United States using the best available information for the year 1992. We95

then used the calibrated model coefficients to predict incremental nitrogen yield for 1992,96

2001, and two other hypothetical scenarios: one where we set the hydrologic conditions to97

1992 levels and source contributions to 2001 levels, and a second where we set hydrologic98

conditions to 2001 levels and source contributions to 1992 levels. For model parameters99

that we assumed were constant over the period of analysis (soil permeability and drainage100

density), we used the base data available as part of the 2.8 version of the SPARROW101

model [Schwarz et al., 2006]. For time dependent variables (e.g., observed loading, ap-102



plication rates, and land use conditions), we determined the state conditions for the two103

study years. Our analysis focused on the incremental total nitrogen yield estimated by104

the model for over 60,000 river reaches in the contiguous United States. Incremental yield105

is defined in the SPARROW model as the total flux delivered from the incremental water-106

shed to the reach, normalized by the watershed drainage area with units of kg ha−1 yr−1.107

The model output quantifies nitrogen yield for each nitrogen source considered within the108

model. A key assumption of our study was to neglect human management practices meant109

to reduce nitrogen loading (e.g., reductions wastewater point source loadings and the use110

of Best Management Practices (BMPs) to control nitrogen runoff). This assumption was111

necessary because of a lack of data for quantifying the impact of these activities, however112

the results of the analysis should be interpreted in light of this key model limitation.113

In the following section we provide details of the study methodology summarized in114

the previous paragraph. This section is followed by a presentation and discussion of115

the study results as a means for understanding changes in the spatial distribution of116

incremental nitrogen yield over the decadal study period. Finally we present a summary117

and concluding statements from this work, along with suggested extensions to the study118

materials and methodology that could be accomplished through future research.119

2. Materials and Methods

2.1. Model Description

SPARROW is a non-linear regression model that relates measured loadings at moni-120

toring stations to point and nonpoint source loadings, waterbody properties, and water-121

shed attributes in order to predict long-term mean annual instream load for unmonitored122

reaches [Schwarz et al., 2006]. The measured instream loadings serve as the dependent123



variable in the regression, while point and nonpoint sources, waterbody properties, and124

watershed attributes serve as the independent variables. Nitrogen source terms consid-125

ered by the model include direct sources such as population related sources, atmospheric126

deposition, fertilizer application and animal waste, as well as indirect sources such as127

non-agricultural land use. Watershed attributes typically considered in SPARROW in-128

clude precipitation, temperature, soil permeability and stream density, while waterbody129

attributes include flow rate, velocity and hydraulic loading (for lakes and reservoirs).130

The SPARROW model formulation used in this study states that the mean annual total131

nitrogen loading observed at reach i is132

Li =[{ ∑

j∈J(i)

Lj}A(Zs

i , Zri ;κs, κr) +

{ N∑n=1

Sn,iβn}Dn(ZD

i ;α)A′(Zsi , Z

ri ;κs, κr)

]εi (1)

where the first summation term is the amount of flux that leaves each of the adjacent up-133

stream reaches for reach i where Lj represents the measured or estimated flux leaving the134

adjacent upstream reach j. The function A(.) represents the loss of mass due to transport135

to the downstream node of reach i (Figure 1). The vectors Zsi and Zr

i are measured stream136

and reservoir characteristics, while κs and κr are corresponding coefficient vectors. If the137

waterbody i is a stream, then Zsi and κs define the function A(.) and if the waterbody i138

is a lake or reservoir, then Zri and κr define the function A(.).139

The second summation term is the amount of mass originating within the watershed140

that is delivered to the waterbody i. N is the total number of mass sources, n is an141

individual mass source, and Sn,i is the contribution from mass source n in reach i. βn is a142

regression coefficient that represents an array of source-specific coefficients and serves as143



a conversion factor between the source units and flux units estimated by the model. The144

function D(.) represents the land-to-water delivery process where α is the estimated vector145

of coefficients and ZDi is the vector of watershed attributes. The term A′(.) represents146

the instream mass loss for reach i. Finally, εi is the multiplicative error term defined147

by the model, which is independent and identically distributed across watersheds in the148

intervening drainage area between monitoring stations [Alexander and Smith, 2005; Hoos149

and McMahon, 2009].150

Land-to-water delivery is represented in our model by Equation 2 that represents a151

first-order decay process152

Dn(ZDi ;α) = exp(−α′ZD

i ) (2)

where α′ is an array of the model coefficients that describe the land-to-water delivery153

for each element in the watershed attribute array, ZDi . Instream transport is represented154

within the model by Equation 3 that models loss as a first-order decay process155

A(Zsi , Z

ri ;κs, κr) = exp(−κs′T i,j) (3)

where κs′ is the array of decay coefficients for streams classified by their mean annual flow156

rates. The decay coefficients are estimations of mass loss per unit stream length. Ti,j is157

an array of waterbody characteristics for the flow path. The first-order decay is modeling158

losses due to physical processes occurring within the stream such as denitrification and159

sedimentation [Alexander and Smith, 2005]. Finally, for reaches that represent lakes or160

reservoirs, A(Zsi , Z

ri ;κs, κr) takes the form of Equation 4 that models nitrogen loss as a161

settling rate162

A(Zsi , Z

ri ;κs, κr) =

(1 +

κrqri

)−1

(4)



where κr is the reservoir decay coefficient estimated by the model and qri is the areal163

hydraulic loading (ratio of reservoir outflow to surface area, in units of distance per time).164

The model is first calibrated using Equations 1-4 to estimate nitrogen loading for all165

monitored reaches. The result of the regression is estimates of the coefficient values for166

β, α, and κ that minimize errors between observed and predicted loadings for monitored167

reaches. Once these coefficients have been determined, the model is applied to predict168

nitrogen delivery for unmonitored reaches within the river network. The model results169

in predictions of incremental and total nitrogen yield for each reach in the river network170

dataset and for each nitrogen source considered by the model. While there are limitations171

to the modeling approach used by SPARROW, as we will discuss in greater detail in172

Section 3.3, we have used the model because it offers a practical blend of process-based173

and empirical modeling appropriate for regional-scale assessments, where it is difficult to174

parameterize physical-based models of hydrologic and biogeochemical processes.175

In this study we considered four model scenarios: Model I (1992), Model II (2001176

Hydrology), Model III (2001 Sources) and Model IV (2001). Model I represents the177

1992 conditions. All the sources, land-to-water delivery terms, and mean annual loadings178

represent the year 1992. Model II represents Model I modified so that the precipitation179

and mean air temperature are set to 2001 conditions. We use this model scenario to180

estimate how the hydrologic changes (precipitation and evaporation, which is related181

to air temperature) impacted nitrogen delivery. Model III represents Model I modified182

so that the source variables are set to 2001 conditions. This model scenario allows us to183

estimate loading changes due to changes in nitrogen sources. Finally, Model IV represents184

the 2001 scenario where all the sources and land-to-water delivery terms represent the year185



2001. Our initial approach was to calibrate the model to 1992 and 2001 observed loading186

data separately, but there were insufficient instream concentration observations available187

for 2001 to support this approach. Therefore we calibrated the SPARROW model for188

1992 conditions and then simulated stream loads for 2001 conditions assuming that the189

SPARROW model coefficients were unchanged and only state conditions changed. We190

then used the SPARROW model coefficients from the Model I calibration to also predict191

loadings for the Model II and Model III scenarios. Our justification for this assumption192

is that because there were no major departures in state conditions between the decade193

separating the two study periods, we can assume that the statistical relationship that194

forms the basis of the model formulation is valid for both years. To verify our assumptions,195

we evaluated the 2001 model by comparing predicted loadings to the limited set of available196

observed loadings. A similar approach of calibrating the model for 1992 and then using197

the calibration coefficients to predict for 2002 was used by Alexander et al. [2008] for the198

Mississippi River Basin.199

Once the model has been calibrated and predicted, the incremental yield for each catch-200

ment was estimated as201

Y ieldi ={ N∑n=1

Sn,iβn}×

Dn(ZDi ;α)A′(Zs

i , Zri ;κs, κr)

/Areai (5)

where i is the number of incremental catchments, Areai is the area of the catchment and202

Y ieldi is the incremental yield in kg ha−1 yr−1. Incremental yield gives the yield estimates203

that originated in that specific catchment without considering the upstream contributions.204

The incremental yield model output was the primarily focus of this study because we were205



interested in how loads were delivered to streams under different hydrologic and sourcing206

conditions.207

2.2. Data Preparation

It is typical to use the Enhanced River Reach File 1 (ERF1) dataset as the river network208

representation for continental-scale SPARROW applications [Nolan et al., 2002]. ERF1209

is a digital stream network at a 1:500,000 spatial scale for the contiguous United States210

that is an improvement over the earlier River Reach File 1 (RF1) dataset [DeWald et al.,211

1985; Alexander et al., 1999]. The stream network consists of more than 60,000 reaches212

with mean reach length of 17 km and total reach length of 1 million km. Approximately213

2,000 of the river reach features represent large reservoirs with a capacity greater than 6214

million m3 [Smith et al., 1997]. Mean streamflow, stream velocity, and time of travel are215

included as reach attributes in the dataset and used to model instream loss rates due to216

sedimentation and denitrification in the SPARROW model [DeWald et al., 1985]. The217

dataset also includes attributes of stream morphology and hydraulic properties, for exam-218

ple incremental and total drainage area, drainage density, mean water depth, and areal219

hydraulic load. Mean stream velocity was estimated using a regression based approach220

that relates stream velocity to long term mean streamflow and stream order. Travel time221

was then estimated by dividing the reach channel length by the mean stream velocity222

[DeWald et al., 1985].223

Watersheds for each reach within the ERF1 dataset were derived using terrain pro-224

cessing algorithms and the Hydro 1K Digital Elevation Model [U.S. Geological Survey ,225

2000]. Watershed attributes considered in previous applications of SPARROW include226

precipitation, long term average streamflow, incremental drainage area, soil permeability,227



slope, drainage density, reach length and mean air temperature [Smith et al., 1997]. How-228

ever, this previous work found that only soil permeability, drainage density and mean air229

temperature were statistically significant for explaining the nitrogen fate and transport at230

the regional spatial scale [Smith et al., 1997]. Later work has shown precipitation to also231

be a statistically significant watershed attribute in nitrogen modeling using SPARROW232

[Hoos and McMahon, 2009]. Therefore, we selected soil permeability, drainage density,233

mean air temperature, and precipitation as watershed attributes in our model formula-234

tion. Soil permeability was estimated for each watershed using the State Soil Geographic235

(STATSGO) database [Schwarz and Alexander , 1995]. Drainage density is defined as the236

the ratio of stream length to the drainage area and was calculated from the ERF1 reach237

length and the area of the Hydro 1K-derived watershed associated with that reach.238

We used PRISM (Parameter-elevation Regressions on Independent Slopes Model) data239

[PRISM , 2004] for temperature and precipitation estimates. After creating the maximum240

and minimum temperature grid from the data available in PRISM, we averaged the max-241

imum and minimum grids to estimate the average temperature for the years 1992 and242

2001. After creating the average temperature grid, we estimated temperatures for each243

ERF1 reach catchment (Figure 4). We created the precipitation grid for the year 1992 as244

the average from the PRISM data for the years 1991, 1992, and 1993. We followed the245

same procedure to estimate the average precipitation for the year 2001 by using the years246

2000, 2001, and 2002. We took this approach to dampen the variability that can occur247

in annual data. PRISM precipitation data indicated that 2001 was dryer than 1992 for248

most of the United States, especially for the West, Midwest and Southwest (Figure 4).249



Total nitrogen loading, the dependent variable in the model, was quantified using data250

from the United States Geological Survey (USGS) National Stream Quality Accounting251

Network (NASQAN). NASQAN was established by the USGS in 1974 to provide a long-252

term, systematically collected baseline water chemistry dataset for the nation [Ficke and253

Hawkinson, 1975; Alexander et al., 1996]. We define Total Nitrogen (TN) in this study254

as nitrate, nitrite, and total Kjeldahl nitrogen in unfiltered samples. Annual loading was255

estimated by using the Fluxmaster program [Schwarz et al., 2006]. Fluxmaster predicts256

the continuous daily load from continuous daily observed streamflow and discontinuous257

nitrogen concentration data by using a nonlinear regression model, and then calculates258

annual loading by taking averages of the daily estimates of total nitrogen loading. In the259

estimation process we used stations that have at least 15 observations to reduce estimation260

uncertainty.261

We selected the following Fluxmaster model to relate measured nitrogen concentration262

to streamflow and other explanatory variables.263

ln(l) = λ0 + λ1t+ λ2sin(2πt)

+λ3cos(2πt) + λ4ln(q) + λ5[ln(q)]2 + α (6)

where l is the instantaneous nitrogen transport, t is decimal time to account for temporal264

trends [Robertson et al., 2006], q is instantaneous discharge and λ0 through λ5 are regres-265

sion coefficients. The term α is the sampling and model error assumed to be independent266

and identically distributed, while the trigonometric terms approximate seasonal variations267

in transport. The mean annual loading was calculated as268

L =1

365

365∑i=1

exp[λ0 + λ1ti + λ2sin(2πti)



+λ3cos(2πti) + λ4ln(qi) + λ5[ln(qi)]2]Vf (7)

where ti is the ith day of the base year in decimal format, qi is the average streamflow269

of the ith day of the year over the multi-year period of the streamflow data, and Vf is270

the minimum variance bias re-transformation correction factor. The minimum variance271

unbiased estimator procedure [Cohn et al., 1989] was used in the model. The time period272

used for estimating the average annual loads was 1970 to 2006, as described in the following273

paragraph, although for many stations the availability of data was only 1970 to 1995 due274

to budget reductions and the resulting discontinuation of monitoring for certain stations275

in 1995.276

We obtained total nitrogen concentration, instantaneous streamflow, and daily average277

streamflow observations for NASQAN sites for the time period 1970-2008. We auto-278

mated the data retrieval process by using web services provided through the Consortium279

of Universities for the Advancement of Hydrologic Science, Inc. (CUAHSI) Hydrologic280

Information System (HIS) [Maidment , 2008; Goodall et al., 2008]. We used nitrogen con-281

centration and daily streamflow data from 1970 to 2000 and detrended the water quality282

and flow model in Fluxmaster to 1992 to estimate long term mean loading. Using this283

time period, we were able to estimate loading for 354 monitoring stations for our 1992284

analysis. We also estimated loading for 2001 with a flow-concentration relationship more285

closely targeted to the study year using nitrogen concentration and daily streamflow data286

from 1996 to 2006 and then detrended to 2001. In this way we were able to estimate287

loading for 122 stations. Again, this 2001 loading dataset was not used for calibration of288

the SPARROW model because there were too few stations, but instead for comparison289



of the estimated loading for 2001 from monitoring stations with the predicted loading for290

2001 from the SPARROW Model IV simulation.291

Nitrogen sources considered in the model are population related sources, atmospheric de-292

position, fertilizer application, livestock waste, and non-agricultural land. Using publicly293

available datasets, we quantified each source for the years 1992 and 2001. We used human294

population estimates for the population related sources, similar to previous SPARROW295

studies [Alexander et al., 2008]. The assumption is that contribution from population296

related sources like wastewater effluent, municipal waste, and urban runoff are related to297

human population. We used county level population estimates from the United States298

Census Bureau [U.S. Bureau of Census , 2010] for 1992 and 2001 and ArcGIS R© to calcu-299

late the population density for the contiguous United States at 1km resolution. We then300

estimated the population count for each ERF1 watershed for the years 1992 and 2001.301

Atmospheric deposition is a well known and important source of nitrogen, specifically302

nitrate, to streams [Jones et al., 2001]. We considered wet deposition of inorganic nitro-303

gen (nitrate and ammonia) (kg yr−1) as a measure of atmospheric deposition following304

the approach used in previous SPARROW model applications [Smith et al., 1997]. We305

did not include ammonia deposition to avoid double counting of agricultural nitrogen306

input [Howarth et al., 1996]. Annual estimates from the National Atmospheric Deposi-307

tion Program [NADP , 2010] were used to estimate wet deposition of inorganic nitrogen.308

These monitoring station estimates were converted to a continuous grid of 1km resolution309

using an inverse-distance weighting interpolation method. Similar to precipitation data,310

we averaged 1991, 1992, and 1993 to create the 1992 deposition input dataset. We then311

summarized the atmospheric deposition loadings for each watershed in the study region.312



We followed the same procedure for the 2001 nitrogen deposition, averaging deposition313

estimates for 2000, 2001, and 2002.314

Fertilizer application to both agricultural lands and to urban lands for lawn maintenance315

is another major source of instream nitrogen. The Association of American Plant Food316

Control Officials (AAPFCO) [Ruddy et al., 2006] used state total sales rather than county317

level sales to estimate fertilizer application because county level fertilizer sales data are318

not reliable. Sometimes multi-county distributors report their sales for a single county319

rather than the actual counties. Moreover farmers can buy fertilizer from one county and320

use it in another county, so the spatial distribution of fertilizer sales may be inaccurately321

represented. Data on the state-level annual sales of commercially produced fertilizer from322

American Plant Food Control Officials (AAPFCO) and data on the county-level fertilizer323

expenditure that were obtained from the Census of Agriculture were used to estimate324

county level nitrogen input from farm fertilizer use in kilograms of nitrogen [Ruddy et al.,325

2006]. State level non-farm fertilizer sales were estimated from the population data and326

converted to nitrogen inputs in kilograms of nitrogen [Ruddy et al., 2006]. This county327

level nitrogen input was then summarized to SPARROW watershed level estimates at 1328

km resolution for both 1992 and 2001.329

Finally, livestock waste was considered as another source of instream nitrogen loading in330

our model. Nitrogen input from livestock waste was estimated from county-level livestock331

population data collected by the Census of Agriculture. Ruddy et al. [2006] presented332

a county level estimate of nitrogen in the livestock waste from confined and unconfined333

animals. Both recoverable manure from confined animals and unrecoverable manure from334

confined and unconfined animals were included in this estimation. We estimated SPAR-335



ROW watershed level nitrogen input from the county level dataset for 1992 and 2001 at336

1km resolution.337

We used the National Land Cover Dataset (NLCD) 1992/2001 Retrofit Land Cover338

Change Product (Fry et all., 2009) to assess land use in the years 1992 and 2001. It339

is not possible to directly use the NLCD 1992 and 2001 land cover products because340

the products were generated using different classification methodologies, source image341

seasonality, georegistration approaches, mapping methodologies, and land use classes [U.S.342

Geological Survey , 2001]. We instead reconstructed two land use datasets from the recently343

produced NLCD Land Cover Change Product that include major land use types of urban,344

forest, crop land, grass land and non-agricultural land. Urban land includes areas of low,345

medium, and high intensity development with a mixture of constructed materials and346

vegetation [U.S. Geological Survey , 2001]. Forest land includes deciduous forest, evergreen347

forest, and mixed forest; Cropland includes cultivated crops and pasture lands; Pasture348

lands include both grasses and legumes for livestock grazing; Grassland includes both349

grassland and shrubs; Non-agricultural land represents the combination of urban, forest,350

and grasslands. Land use grids for the two study years were used to estimate the total351

area of each land use type, for each year, and for each of the watersheds.352

We used the data described in the previous paragraphs to construct the inputs for the353

four previously described model scenarios. For Model I, we both calibrated the model354

and used the calibrated model to predict 1992 loadings. For the remaining three model355

scenarios, we used the parameters from the Model I calibration to predict loadings. We356

used the 2.8 version of SPARROW, the latest version of the model at the time of this357

study.358



3. Results and Discussion

3.1. Model Calibration

Model I (1992) resulted in R2 values for predicted log of flux of 0.885 and R2 values for359

predicted log of yield of 0.802 (Table 1, Figure 2). These R2 values are similar to those360

obtained in previous SPARROW model applications. For example, Smith et al. [1997]361

reported results from a national SPARROW model for total nitrogen using 414 NASQAN362

stations in which their R2 values for log of flux were 0.874 and Hoos and McMahon [2009]363

reported R2 values for flux and yield were 0.96 and 0.68, respectively for an application of364

SPARROW for the Southeastern United States. The model residuals (Figure 2) showed365

some signs of a spatial bias with the highest and lowest loads corresponding to the largest366

and smallest rivers, suggesting an over-prediction for larger rivers and an under-prediction367

for small rivers. Some possible bias was also present in specific regions such as the Pacific368

Northwest, which showed over-predictions in general, and the Midwest, which showed369

under-predictions in general. SPARROW predicted loading for Model IV (2001) was also370

compared to actual loading observations estimated using the Fluxmaster program based371

on the flow and concentration data available from 1996 to 2006 for 122 stations (Figure372

3). The R2 value for this predicted log of flux vs actual log of flux was 0.890 for these373

stations. It should be noted that the high R2 value for the 2001 evaluation may be due in374

part to the 2001 dataset consisting of primarily larger streams when compared to the 1992375

dataset, or to a time lag introduced by using long term flow and concentration data for376

model estimation. Nonetheless, we argue that the evaluation of the model against 2001377

observed loads provides a reasonable level of confidence that the model is able to predict378

loads in this year based on calibrated model coefficients for the 1992 model.379



The SPARROW model includes three coefficients, α, β and κ, that are fit during the380

model calibration process (Table 2). The coefficients have physical meaning in that they381

allow one to understand losses due to land-to-river and instream transport processes.382

The α coefficient results for soil permeability, drainage density, mean annual air tem-383

perature, and precipitation were consistent with expectations and previous SPARROW384

results [Smith et al., 1997; Alexander et al., 2008; Hoos and McMahon, 2009]. All α coeffi-385

cients were statistically significant (p < 0.05). The magnitude of the α coefficients for the386

different watershed attributes provides important information on how soil permeability,387

drainage density, mean annual air temperature and precipitation influence the efficiency388

with which nitrogen is delivered from the land to waterbodies. As expected, the coeffi-389

cients for soil permeability and temperature were negative. Highly permeable soils will390

have higher absorption capacities for nutrients, which act to resist nitrogen transport to391

streams [Smith et al., 1997]. Temperature has a negative correlation with nitrogen trans-392

port because increased temperature results in an increased denitrification rate, therefore393

decreasing the proportion of nitrogen transported to waterbodies [Smith et al., 1997].394

Drainage density, the ratio of stream length to drainage area for a watershed, is positively395

correlated with nitrogen delivery because having a higher drainage density increases the396

ability for nitrogen to be delivered from the landscape to waterbodies. Similarly, precip-397

itation is positively correlated with nitrogen transport as higher precipitation indicates398

higher runoff.399

For direct nitrogen sources, the β coefficients account for possible variation in source es-400

timates and will vary between sources [Smith et al., 1997]. For indirect nitrogen sources,401

the β coefficients represent a source term and provide information about the quantity402



of nitrogen originating from different sources. The β coefficient for population related403

sources was 3.57 kg person−1 yr−1 and was statistically significant (p < 0.05). The β coef-404

ficients for atmospheric deposition, fertilizer application, and livestock waste production405

were 0.28, 0.25, and 0.08, respectively, and the β coefficient for non-agricultural land was406

310 kg km−2 yr−1 . The β coefficients for fertilizer application and non-agricultural land407

were statistically significant (p < 0.05), but the β coefficients for atmospheric deposition408

and livestock waste production were not statistically significant (p > 0.05). Land use409

types urban, crop, forest, and grass land were used as variables in the preliminary model410

simulations but later dropped based on the following criteria. After considering statistical411

significance, if a variable was not statistically significant, we used variance inflation factor412

(VIF) to determine if the multicollinearity was responsible for this. We also used eigen-413

spread values to identify multicollinearity. If the eigenspread value was greater than 100,414

we dropped that predictor variable from the model [Schwarz et al., 2006]. Our final selec-415

tion of the predictor variables was based on constant variance and minimum correlation.416

Cropland was dropped because of the strong collinearity between cropland and fertilizer417

application. For the same reason only the population related source was included and the418

urban land source was dropped from the final model.419

The model calibration resulted in instream loss coefficients (κ) that were greater for low-420

flow streams (Q < 28.3 m3 s−1) compared to medium-flow streams (28.3 m3 s−1 ≤ Q ≤ 283421

m3 s−1). These coefficients were found to be statistically significant for low flow streams422

(p < 0.05), but not for medium streams (p > 0.05). This is consistent with previous423

SPARROW model results and the conclusion that loss rate decreases with increasing424

stream size [Alexander et al., 2000; Hoos and McMahon, 2009]. Model prediction for425



the reservoir loss coefficient (κr) was 7.18 m yr−1 in 1992, was statistically significant (p426

< 0.05), and was consistent with Hoos and McMahon [2009], which reported a slightly427

higher reservoir loss rate of 13.1 m yr−1 for the Southeastern United States. The standard428

errors for the coefficients (Table 2) provide a measure of the confidence intervals of the429

coefficient estimates. The standard error values are in line with prior SPARROW studies430

and could be used through future work to quantify uncertainty of the model estimates431

reported in this study.432

3.2. Model Predictions

When comparing overall change in loading between Model I (1992) and Model IV (2001),433

the results indicate a decrease in the median incremental nitrogen yield of 0.67 kg ha−1434

yr−1 (or 8.21%) from 8.16 kg ha−1 yr−1 in 1992 to 7.49 kg ha−1 yr−1 in 2001 (Table 3).435

Table 3 presents various statistical quantities for the predictions including the standard436

deviation for predicted incremental total nitrogen yield (kg ha−1 yr−1) that provide a437

measure of the distribution of yield results across the entire study area. We have chosen438

to summarize the incremental nitrogen yield prediction results using the median value439

because it is less influenced by very high incremental yields predicted for larger river440

basin systems. That said, we acknowledge that because first-order streams dominate441

the stream network, the median will be weighted toward changes in smaller streams.442

Alexander et al. [2008] also indicated a slight overall decrease in both simulated loadings443

based on a SPARROW model, as well as monitoring-based loadings of nitrogen in streams444

over a similar time period for the Mississippi River Basin. Comparing Model I vs Model445

II (2001 Hydrology) and then Model I vs Model III (2001 Sources) reveals that this446

decrease was mainly due to hydrologic differences rather than variability in nitrogen source447



input. The estimates of median incremental nitrogen yield between Model II and Model448

I indicate a decrease of 0.71 kg ha−1 yr−1 (or 8.70%). This decrease is the result of449

changes in precipitation and mean annual air temperature between 1992 to 2001. The450

precipitation map [PRISM , 2004] indicated a decrease in precipitation in 2001 compared451

to 1992 (Figure 4), and precipitation is an important nitrogen delivery variable for land-452

to-water transport. Comparing estimates between Model III and Model I indicates an453

increase of 0.01 kg ha−1 yr−1 (or 0.12%) in incremental nitrogen yield (Table 3). This454

result suggests that only taking into account changes in sources does not account for the455

changes in incremental nitrogen yield over the period of study. Previous studies in the456

Mississippi River Basin also showed an increase in nitrogen sources and loading to streams457

before the early 1980s, but after that time no significant trend was observed [Goolsby et al.,458

1999; National Agricultural Statistics Service (NASS), 1998; Alexander and Smith, 1990;459

Council of Environmental Quality (CEQ), 1989].460

The total nitrogen yield map (Figure 5a) shows the overall incremental yield scenario461

for 1992 (Model I). This figure shows that the incremental yield estimates are higher in the462

Upper Mississippi, the Ohio, the southeastern portion of the Missouri, the Tennessee and463

the Lower Mississippi Basins, but also in the Northeastern and the Pacific West Coast464

regions. A common characteristic of these regions is that they have either significant465

agricultural lands or higher human populations, which account for the higher incremental466

yields. Alexander et al. [2008] showed a similar pattern of nitrogen yield for the Mississippi467

River Basin. The results shown in Figure 5a also suggest that the incremental nitrogen468

yield estimates are highest in the wettest areas of the country and lowest in the driest469

areas. Nitrogen can be more easily transported in wetter climates because of higher470



precipitation, runoff, and discharge. On the other hand, in arid and semiarid climates,471

low amounts of precipitation and high evaporation rates result in limited runoff and, as a472

result, reduced nitrogen yields. This may be the reason for the lower nitrogen yield from473

large populated cities in the Southwestern United States.474

Figure 5b shows differences in the incremental yield scenario if hydrologic inputs (mean475

annual air temperature and precipitation) are set to 2001 conditions but nitrogen source476

data is held at 1992 conditions (Model I vs Model II). Analyzing the input data shows477

that the change in precipitation map has a similar pattern as seen in 5b, indicating478

the importance of precipitation in particular in nitrogen delivery due to nonpoint source479

pollution transport. Figure 5c shows the results of a scenario where we used 2001 source480

contribution but 1992 hydrologic conditions to quantify how change in sources only would481

affect incremental nitrogen yield (Model I vs Model III). Comparing Figure 5c with Figure482

5a suggests that the region in the Mississippi Basin with the highest nitrogen yield in 1992483

actually saw a decrease in yield when considering only changes in nitrogen sources and484

controlling for hydrologic/climate changes. Finally Figure 5d presents the overall change485

in incremental yield during the period 1992 to 2001 (Model I and Model IV). This map486

shows that some regions of the contiguous United States produced more incremental487

nitrogen yield, despite the fact that the median yield decreased during the period of488

analysis. Furthermore, by comparing Figures 5b and 5c to Figure 5d, we can observe489

changes in incremental nitrogen yield over the study period and determine whether they490

were due to variability in nitrogen sources or changes in hydrologic/climate conditions.491

For example, the increase in the Ohio and Tennessee Basins appears to be primarily due492



to hydrologic changes, while the increase in the northern portion of the Missouri Basin493

appears to be primarily due to increases in nitrogen sources.494

Estimates of median incremental yield for each model scenario, separated by each ni-495

trogen source, are presented in Figure 6. Comparing Model I to Model IV provides496

an estimate of median incremental yield difference between 1992 and 2001 and indi-497

cates a decrease of 0.01 kg ha−1 yr−1(2.8%) for population related sources, 0.14 kg ha−1498

yr−1(15%) for atmospheric deposition, 0.08 kg ha−1 yr−1(5.4%) for fertilizer application,499

0.03 kg ha−1 yr−1(6.1%) for livestock waste production, and 0.22 kg ha−1 yr−1(8.0%) for500

non-agricultural lands. Comparison of Model I vs Model III results, which controls for501

hydrologic changes, provides insight into how changes in nitrogen sources alone during502

the study period impacted nitrogen yield to streams. The results show that population503

related sources had the largest percent increase in nitrogen yield (8.4%), fertilizer applica-504

tion had the second largest percent increase (2.9%), and livestock waste had the smallest505

(0.7%) increase. Median incremental nitrogen yield from atmospheric deposition and non-506

agricultural land decreased by 6.6% and 0.15%, respectively. The decrease in nitrogen507

yield from atmospheric deposition was likely due to the fact that we used wet deposition508

as an input dataset and, by doing so, this source already accounts for hydrologic changes.509

Controlling for variability in nitrogen sources (Model I vs Model II) showed population510

related sources had the largest percent decrease (10%) in nitrogen yield. Thus, while yield511

from population related sources had an overall decrease of 2.8% when comparing Model512

I and Model IV, considering also the results from Models II and III suggests that this513

decrease was primarily due to 2001 being a drier year than 1992. Finally, we noted that514

the percent yield decrease when controlling for sources (Model I vs Model II) was similar515



for all sources (ranging between a 7% and 10% decrease for the different nitrogen sources516

considered). When interpreting these results, it should be noted that even though human517

population is used as a surrogate for point source and non point urban sources, population518

is diffuse in nature and the relationship between changes in point sources and changes in519

population is not ideal but necessary given available information. It should also be noted520

that fertilizer application and livestock waste were assigned to reaches based on county521

level data, and this data might exaggerate these sources for many counties outside of the522

agricultural Midwest. Also, as mentioned previously, management practices intended to523

reduce nitrogen in waters such as improvements in wastewater treatment, construction of524

urban retention ponds as Best Management Practices (BMPs), preserving riparian buffer525

zones, and reductions in automobile emission (if not reflected in NADP measurements) are526

not considered in this study because we are unaware of accurate national-scale datasets527

for quantifying the impacts of these activities.528

The spatial distribution of changes in yield for each source when comparing Model529

I vs Model IV suggests that nitrogen yields from population related sources increased530

in the western United States, Texas, and Florida (Figure 7). However, it also suggests531

an increase in the Tennessee and Ohio Basins. This is the same region that showed532

higher precipitation rates between 2001 and 1992, which may explain the higher yield533

from population sources as precipitation may have caused an increase in nonpoint source534

pollution from populated areas in this region. Nitrogen yields from atmospheric deposition535

show a pattern nearly identical to the input dataset of changes in nitrogen deposition536

over the study period. These data show widespread decreases in wet nitrate deposition537

in the Midwest and Northeast over the period of analysis [NADP , 2010]. Fertilizer yield538



increased in the West, upper portion of the Missouri Basin, Ohio Basin, and the Northeast,539

while it decreased in the Southeast and Upper Mississippi Basin. Alexander et al. [2008]540

found that the fertilizer sources decreased in the Upper Mississippi Basin and increased541

in the Missouri Basin due to expanded corn and soybean production, which agree with542

our findings. However, Alexander et al. [2008] found a decrease in fertilizer application543

for the Ohio Basin, but we found an increase for a portion of this Basin in our study.544

The increases due to fertilizer in the Ohio Basin and the upper portion of the Missouri545

Basin follow the pattern shown in the overall incremental yield change map for 1992 to546

2001 (Figure 5d), suggesting that fertilizer increases were the primary source of overall547

nitrogen yield increase in these regions. Increases due to livestock waste showed a scattered548

pattern overall with decreases in the lower portion of the Mississippi Basin. Changes due549

to the non-agricultural land followed the same pattern as precipitation changes because550

non-agriculture land is well distributed across the study region.551

Figure 8 presents source share results for each nitrogen source considered in the model.552

Source share is defined as the incremental yield for a specific source of nitrogen divided553

by the total incremental yield for a given watershed. Given the definition of incremental554

yield (Equation 5), it can be shown that the watershed property term (ZD) and the in-555

stream transport terms (Zs and Zr) cancel out when calculating source share. In other556

words, because the transport factors are applied equally to all sources, the changes in cli-557

mate/hydrology transport factors do not impact the incremental measures of the source558

shares, which are applied equally for all sources. Therefore, Model I source share results559

match Model II source share results and, similarly, Model III source share results match560

Model IV source share results. We presented Model I and Model IV source share re-561



sults in Figure 8. The model results show an increase in source share contribution from562

population related sources and livestock waste, and a decrease in source share from at-563

mospheric deposition, fertilizer application, and non-agricultural land. Over the period of564

analysis population increased by approximately 8% in the contiguous United States [U.S.565

Bureau of Census , 2010] and population related source share increased by 11.5%. The566

source share for atmospheric deposition decreased by 6.17%, however this decrease may567

be related to lower precipitation in 2001 compared to 1992, as the model considered wet568

deposition of inorganic nitrogen (nitrate and ammonia) as an estimate for atmospheric569

deposition. The decrease in source share from fertilizer application was 3.0%. While the570

fertilizer application rate data from USDA indicated increases in fertilizer use for some571

regions of the United States, the model results suggest an overall decrease in source share572

contribution for fertilizer. This decrease in fertilizer source share could simply be the573

result of an increase in source share contribution from population related sources. How-574

ever, this fertilizer source share decrease did not account for changes in farm management575

practices such as conservation tillage, nutrient management, and BMPs, and nitrogen576

removal by increasing crop yields, which all likely contributed to reducing nitrogen yield577

from agricultural lands over the period of analysis. Source share from livestock waste578

production increased by 1.2%, while source share from non-agricultural land decreased by579

1.5%. Both sources showed little change as these sources remain nearly constant over the580

period of analysis.581

To understand the impact of land use type on incremental yield, we estimated incre-582

mental yield loading from watersheds that have a dominant land use type (Figure 9 and583

Table 4). If the land cover of the watershed was more than 90% urban land, we con-584



sidered it as an urban watershed. We did the same for cropland and grass land. If the585

land cover was more than 95% forest land, we considered it to be a forested watershed.586

Results showed that while median incremental yield from urban watersheds experienced587

only a 0.4% decrease from 1992 to 2001 (Model I vs Model IV results), this was due to an588

offset between hydrologic changes and source contribution variability. If we consider only589

changes in hydrology between 1992 and 2001 (Model I vs Model II), the results show a590

decrease of 7.4% in median incremental yield. If we consider changes in source contribu-591

tion only over the same time period (Model I vs Model III ), the results show an increase592

of 6.8%. Urban watersheds showed the most significant increase in nitrogen yield due to593

variability in source contribution after controlling for hydrologic changes, however this594

result should be interpreted in light of the fact that the study does not account for reduc-595

tions in point sources of nitrogen that were implemented over the study period. Results596

also showed an overall decrease in median incremental yield from cropland watersheds597

of 16% from 1992 to 2001 (Model I vs Model IV). However, this decrease was primar-598

ily due to hydrologic changes and not to decrease in source contributions. Incremental599

yield decrease for cropland due to hydrologic changes (Model I vs Model II) was 12%600

while decrease due to source contribution variability was 2.1% (Model I vs. Model III).601

Forested watersheds showed an overall median incremental yield decrease of 5.8% (Model602

I vs Model IV), but this change was primarily due to hydrologic changes (6.3%; Model603

I vs Model II) and not decreases in source contribution (0.5%; Model I vs Model III).604

Finally grassland watersheds showed a 3.8% overall decrease (Model I vs Model IV) with605

a 7% decrease when considering only hydrologic changes (Model I vs Model II) and 3.5%606

increase when considering only source contribution variability (Model I vs Model III) in607



median incremental yield. The median incremental yield from the different land use types608

were similar to previously published values (Table 4; Frink [1991]; Ritter [1988]; Beaulac609

and Reckhow [1982]).610

3.3. Model Limitations and Assumptions

While we have made note of limitations of our model, data, and methodology through-611

out this discussion, we highlight the most significant of these limitations and assumptions612

here. One key limitation is that SPARROW is a semi-empirical model and therefore has613

inherit differences compared to a more mechanistic modeling approach. However, there614

are challenges and issues associated with using a more mechanistic model at a national615

scale including data availability, computational requirements, the need for some empirical616

assumptions even in mechanistic models, subgrid heterogeneity, and problems of param-617

eter estimation and calibration [Beven, 1989; Smith et al., 2008; Oreskes et al., 1994].618

Therefore, while there are advantages to using a more mechanistic model for answering619

our research question, such an approach is not without its own problems and so there620

is still a role to play for simplified versions of process-based models like SPARROW, a621

widely applied model both in the United States and abroad [Elliott et al., 2005; Hoos622

and McMahon, 2009]. While we are on one hand making an argument for SPARROW623

as an appropriate model for our research question, we acknowledge the possibility that624

SPARROW may be overly simplified for addressing our research question because it does625

not, for example, include representations for processes such as nitrogen contributions from626

groundwater, which may be significant for many regions of the country [Ator and Ferrari ,627

1997]. Also previous studies suggest potential problems with the statistical approaches628

used in SPARROW. For example, Qian et al. [2005] suggested a structural weakness in629



the SPARROW model that may cause spatial autocorrelation. Similarly, the fitted SPAR-630

ROW model may be biased because it uses statistical inference of a nonlinear regression631

based on the normality assumption [Fuller , 1995].632

In this study we assumed that some watershed properties and river conditions were633

constant from 1992 to 2001 including soil permeability, drainage density, streamflow rates,634

and stream velocities. Streamflow rates and stream velocities are estimated based on long635

term average condition. We therefore do not suspect large changes in these attributes636

over the short time period of analysis, and we believe it to be a justifiable assumption637

that changes in these parameters from 1992 to 2001 will not significantly alter the findings638

reported in this study. Another important limitation of the study was our inability to639

estimate observed nitrogen loading using NASQAN monitoring stations for 2001 at a640

sufficient number of stations in order to calibrate SPARROW. Due to inadequate observed641

nitrogen data, we were only able to estimate loading for 1992, when sufficient data were642

available to estimate flow-concentration relationships. This estimation was based on long643

term flow condition for 1970-2000 and long term mean load detrended to 1992. To address644

the limitation of lack of monitoring data for 2001 we considered an alternative approach645

to only calibrate the SPARROW model for 1992 and then use the calibration coefficients646

to simulate the loading for 2001. This approach assumes that the model coefficients are647

unchanged and only state conditions change. Alexander et al. [2008] presented a similar648

approach for Mississippi River Basin SPARROW model study. We kept streamflows and649

velocities (which are used to estimate travel times) as long term averages in all four650

model scenarios to be consistent with the long term nitrogen loading estimates. Because651

we were primarily interested in incremental nitrogen yield, we assumed that precipitation652



would capture hydrologic changes and that keeping streamflow and velocity as long term653

averages, to be consistent with instream nitrogen loading data, would be justifiable. An654

extension to this study would be to identify nitrogen concentration and flow observations655

from other datasets outside of NASQAN to use in the analysis to obtain better estimates656

of instream nitrogen loading at monitoring stations for both time periods. In this case,657

the streamflow and velocity estimates should also be updated to each of the base years of658

the simulation.659

4. Conclusion

The goal of this research was to better understand how the variability in source contribu-660

tions (anthropogenic and non anthropogenic) and the changes in hydrology/climate affect661

incremental nitrogen yield within the contiguous United States. We used the SPARROW662

model, land use change products from the National Land Cover Dataset (NLCD), other663

source inputs, watershed characteristics and instream nitrogen loading observations to664

quantify these impacts for more than 60,000 watersheds in the contiguous United States.665

We built four model scenarios to isolate these changes: Model I was a simulation of 1992666

conditions, Model II was a modification of Model I where hydrologic inputs (eg. precipita-667

tion and mean air temperature) were set to 2001 conditions, Model III was a modification668

of Model I where source contribution inputs were set to 2001 conditions, and Model IV669

was a simulation of 2001 conditions.670

The results of this study suggest a decrease of 8.2% in median incremental nitrogen671

yield from 1992 to 2001 (Model I vs Model IV). The decrease was 15% for atmospheric672

deposition, 8.0% for non-agricultural land use, 6.1% for livestock waste, 5.4% for fertilizer673

use, and 2.8% for population related sources. If only changes in nitrogen source contri-674



butions were considered (Model I vs Model III), we observe only a small increase (0.1%)675

in median incremental nitrogen yield. However, if only hydrology related changes were676

considered (Model I vs Model II), we observe a decrease of 8.7% in median incremental677

nitrogen yield. Therefore results of this analysis suggest that hydrologic changes – and678

not decreases in nitrogen source contributions – were primarily responsible for changes679

in nitrogen yield over the period of analysis. The results confirm previous research find-680

ings that suggested significant changes in nitrogen sources to the Mississippi River Basin681

were not observed after the early 1980s [Goolsby et al., 1999; National Agricultural Statis-682

tics Service (NASS), 1998; Alexander and Smith, 1990; Council of Environmental Quality683

(CEQ), 1989]. The results also highlight the importance of precipitation and temperature684

changes on regional scale nitrogen transport.685

The model results suggest decreases in incremental nitrogen yield from some of the686

highest yield producing areas (e.g., Upper Mississippi Basin). After separating hydrologic687

and source contributions using the model scenarios we found that, although some of this688

reduction was due to hydrologic differences between the two years (e.g., 2001 was a drier689

year than 1992), the change was also due to reductions in source contributions, particularly690

in the Mississippi Basin. The model results also show some areas that experienced an691

increase in incremental nitrogen yield over the study period. From the model scenarios692

we observe in some regions this increase was due to increases in source contribution (e.g.,693

the upper portion of the Missouri Basin), but for other regions this increase was due694

primarily to differences in hydrology (e.g., the Pacific Northwest).695

We found from the model scenarios how each source was dependent on hydrologic vs696

source contribution changes. For example, although overall median incremental nitrogen697



yield decreased for population related sources (2.8%; Model I vs Model IV), this overall698

decrease was due to an offset between an increase in source contribution (8.4%; Model I699

vs Model III) and a decrease due to hydrologic changes (10%; Model I vs Model II). When700

incremental nitrogen yield for each source is viewed spatially, we found that changes in701

fertilizer application in particular was responsible for the overall decrease in nitrogen yield702

for the Upper Mississippi Basin and the overall increase in nitrogen yield for the upper703

portion of the Missouri Basin.704

We found that source share of the total nitrogen budget for incremental yield increased705

by 11.5% for population related sources and decreased by 6.17% and 3.0% for atmo-706

spheric deposition and fertilizer application, respectively. Source share for livestock waste707

and non-agricultural land remained nearly constant over the period of analysis. By group-708

ing results for watersheds with dominate land use types, we found that urban watersheds709

showed the largest percent increase in incremental nitrogen yield (6.8%) and cropland710

had largest percent decrease (2.1%), after controlling for hydrologic changes. These re-711

sults suggest that nitrogen from population related sources may becoming a significant712

contributor of incremental nitrogen yield to streams, however it is important to stress713

that this study was not able to account for changes in human management practices over714

the period of analysis that are known to have occur but are not easy to quantify at the715

scale of this study. Another key limitation of our model was that – because of an in-716

sufficient amount of instream nitrogen observation data in and around the year 2001 –717

we were unable to calibrate the model for 2001 conditions. Therefore we assumed that718

model coefficients in SPARROW that describe such properties as instream and land-to-719

water transport were constant between 1992 and 2001. However, we did evaluate the 2001720



model predictions against the available instream nitrogen loading data and model showed721

a good fit to these observed data. One possible extension of this work would be to identify722

other reliable water quality datasets that could be used to improve estimates of instream723

nitrogen loading in 2001 so that the SPARROW model can be re-calibrated for the 2001724

model scenario.725

Acknowledgments. The authors wish to acknowledge Richard Alexander for his as-726

sistance on the paper, in particular with suggestions on the study methodology and help in727

estimating observed loading values using the Fluxmaster program. The authors also wish728

to thank the four anonymous reviewers who provided valuable comments and suggestions729

that greatly improved this study.730

References

Alexander, R., A. Ludtke, K. Fitzgerald, and T. Schertz, Data from Selected U.S. Geolog-731

ical Survey National Stream Water-Quality Monitoring Networks (WQN) on CD-ROM,732

U.S. Geological Survey Open-File Report 96-337, 85 p., 1996.733

Alexander, R. B., and R. A. Smith, County level estimates of nitrogen and phosphorus734

fertilizer use in the United States, 1945 to 1985, U.S. Geological Survey Open-File735

Report 90-130, 1990.736

Alexander, R. B., and R. A. Smith, Moving from monitoring to prediction: The quality737

of the nations streams, available at http://water.usgs.gov/nawqa, 2005.738

Alexander, R. B., J. Brakebill, R. Brew, and R. A. Smith, Erf1 - enhanced river reach file739

1.2, U.S. Geological Survey Open-File Report 99-457, 1999.740



Alexander, R. B., R. A. Smith, and G. E. Schwarz, Effect of stream channel size on the741

delivery of nitrogen to the Gulf of Mexico, Nature, 403 (6771), 758–761, 2000.742

Alexander, R. B., R. A. Smith, G. E. Schwarz, E. W. Boyer, J. V. Nolan, and J. W.743

Brakebill, Differences in phosphorus and nitrogen delivery to the Gulf of Mexico from744

the Mississippi River Basin, Environmental Science & Technology, 42 (3), 822–830, 2008.745

Ator, S. W., and M. J. Ferrari, Nitrate and selected pesticides in ground water of the746

Mid-Atlantic Region, US Geol. Surv.Water- Resources Invest. Report 97-4139, 1997.747

Beaulac, M. N., and K. H. Reckhow, An examination of land use – nutrient export rela-748

tionships, Water Resources Bulletin, 18, 1013–1024, 1982.749

Beven, K., Changing ideas in hydrology – the case of physically-based models, Journal of750

Hydrology, 105 (1-2), 157–172, doi:10.1016/0022-1694(89)90101-7, 1989.751

Boyer, E. W., C. L. Goodale, N. A. Jaworsk, and R. W. Howarth, Anthropogenic ni-752

trogen sources and relationships to riverine nitrogen export in the northeastern USA,753

Biogeochemistry, 57 (1), 137–169, 2002.754

Cohn, T., L. DeLong, E. Gilroy, R. Hirsch, and D. Wells, Estimating constituent loads,755

Water Resources Research, 25 (5), 937–942, 1989.756

Council of Environmental Quality (CEQ), Environmental trends, Executive Office of the757

President, Washington, D.C., 64-65, 1989.758

DeWald, T., R. Horn, R. Greenspun, P. Taylor, L. Manning, and A. Montalbano,759

STORET Reach Retrieval Documentation, U.S. Environmental Protection Agency,760

Washington, D.C., 1985.761

Diaz, R. J., Overview of hypoxia around the world, Journal of Environmental Quality,762

30 (2), 275–281, Annual Meetings of the American Society of Agronomy 1997 Anaheim,763



California, 2001.764

Elliott, A. H., R. B. Alexander, G. E. Schwarz, U. Shankar, J. P. S. Sukias, and G. B.765

McBride, Estimation of nutrient sources and transport for New Zealand using the hybrid766

mechanistic-statistical model SPARROW, 44 (1), 2005.767

Feddema, J. J., K. W. Oleson, G. B. Bonan, L. O. Mearns, L. E. Buja, G. A. Meehl, and768

W. M. Washington, The importance of land-cover change in simulating future climates,769

Science, 310 (5754), 1674–1678, 2005.770

Ficke, J. F., and R. O. Hawkinson, The National Stream Quality Accounting Network771

(NASQAN); some questions and answers, U.S. Geological Survey Circular 719, 23 p.,772

1975.773

Foley, J. A., et al., Global consequences of land use, Science, 309 (5734), 570–574, 2005.774

Frink, C. R., Estimating nutrient exports to estuaries, Journal of Environmental Quality,775

20, 717–724, 1991.776

Fuller, W. A., Estimation in the presence of measurement error, International Statistical777

Review/Revue Internationale de Statistique, 121–141, 1995.778

Galloway, J. N., W. H. Schlesinger, H. Levy, A. Michaels, and J. L. Schnoor, Nitrogen779

fixation: Anthropogenic Enhancement-Environmental Response, Global Biogeochem.780

Cycles, 9 (1), 235–252, 1995.781

Goodall, J. L., J. S. Horsburgh, T. L. Whiteaker, D. R. Maidment, and I. Zaslavsky, A782

first approach to web services for the national water information system, Environmental783

Modelling & Software, 23 (4), 404–411, 2008.784

Goolsby, D. A., W. A. Battaglin, G. B. Lawrence, R. S. Artz, B. T. Aulenbach, D. R. K.785

R. P. Hooper, and G. J. Stensland, Flux and sources of nutrients in the Mississippi-786



Atchafalaya River basin, Report of Task Group 3 to the White House Committee787

on Environment and Natural Resources, Hypoxia Work Group, Federal Register, 64,788

23834–23835, 1999.789

Han, H. J., J. D. Allan, and D. Scavia, Influence of climate and human activities on the790

relationship between watershed nitrogen input and river export, Environmental Science791

& Technology, 43 (6), 1916–1922, 2009.792

Hoos, A. B., and G. McMahon, Spatial analysis of instream nitrogen loads and factors con-793

trolling nitrogen delivery to streams in the Southeastern United States using Spatially794

Referenced Regression on Watershed Attributes (SPARROW) and regional classifica-795

tion frameworks, Hydrological Processes, 23 (16), 2275–2294, 2009.796

Howarth, R. W., et al., Regional nitrogen budgets and riverine N&P fluxes for the797

drainages to the North Atlantic Ocean: Natural and human influences, pp. 75–139,798

1996.799

Jones, K. B., A. C. Neale, M. S. Nash, R. D. V. Remortel, J. D. Wickham, K. H. Ri-800

itters, and R. V. O’Neill, Predicting nutrient and sediment loadings to streams from801

landscape metrics: A multiple watershed study from the United States Mid-Atlantic802

region, Landscape Ecology, 16 (4), 301–312, 2001.803

Juang, J. Y., A. Porporato, P. C. Stoy, M. S. Siqueira, A. C. Oishi, M. Detto, H. S.804

Kim, and G. G. Katuls, Hydrologic and atmospheric controls on initiation of convective805

precipitation events, Water Resources Research, 43 (3), 2007.806

Maidment, D. R., Bringing water data together, Journal of Water Resources Planning807

and Management, 134 (2), 95, doi:10.1061/(ASCE)0733-9496(2008)134:2(95), 2008.808



NADP, National Atmospheric Deposition Program (NRSP-3), NADP Program Office,809

Illinois State Water Survey, 2204 Griffith Dr., Champaign, IL 61820, 2010.810

National Agricultural Statistics Service (NASS), Historic data on livestock, Economic811

and Statistics System, Washington, D.C., on-line data base, 1998.812

Nolan, J., J. Brakebill, R. B. Alexander, and G. Schwarz, Enhanced river reach file 2,813

U.S. Geological Survey Open-File Report 02-40, 2002.814

NRC, Clean coastal waters : understanding and reducing the effects of nutrient pollution,815

National Academy Press, Washington D.C., 2000.816

Oreskes, N., K. Shrader-Frechette, and K. Belitz, Verification, validation, and confir-817

mation of numerical models in the earth sciences, Science, 263 (5147), 641–646, doi:818

10.1126/science.263.5147.641, 1994.819

PRISM, PRISM Climate Group, Oregon State University, available at820

http://www.prismclimate.org, 2004.821

Qian, S., K. Reckhow, J. Zhai, and G. McMahon, Nonlinear regression modeling of nu-822

trient loads in streams: A Bayesian approach, WATER RESOURCES RESEARCH,823

41 (7), doi:10.1029/2005WR003986, 2005.824

Rabalais, N. N., R. E. Turner, and W. J. Wiseman, Gulf of Mexico hypoxia, A.K.A.825

“The Dead Zone”, Annual Review of Ecology and Systematics, 33 (1), 235–263, doi:826

10.1146/annurev.ecolsys.33.010802.150513, 2002.827

Ritter, W. F., Reducing impacts of nonpoint source pollution from agriculture: A review,828

Journal of Environmental Science and Health, 23, 645–667, 1988.829

Roberts, A. D., and S. D. Prince, Effects of urban and non-urban land cover on nitrogen830

and phosphorus runoff to Chesapeake Bay, Ecological Indicators, 10 (2), 459–474, doi:831



10.1016/j.ecolind.2009.07.017, 2010.832

Robertson, D. M., D. A. Saad, and D. M. Heisey, Present and Reference Concentrations833

and Yields of Suspended Sediment in Streams in the Great Lakes Region and Adja-834

cent Areas, U.S. Geological Survey Scientific Investigations Report 2006-5066, 38 pg,835

available at http://pubs.usgs.gov/sir/2006/5066/, 2006.836

Ruddy, B., D. Lorenz, and D. Mueller, County-level estimates of nutrient inputs to the837

land surface of the conterminous United States, 1982–2001, U.S. Geological Survey838

Scientific Investigations Report 2006–5012, 17 p., 2006.839

Schwarz, G. E., and R. B. Alexander, State soil geographic (STATSGO) database for the840

conterminous United States, U.S. Geol. Surv. Open File Rep. 95-449, 1995.841

Schwarz, G. E., A. B. Hoos, R. B. Alexander, and R. A. Smith, The SPARROW surface842

water-quality model: Theory, application, and user documentation, U.S. Geological843

Survey Techniques and Methods Report, Book 6, Chapter B3, 2006.844

Smith, R. A., G. E. Schwarz, and R. B. Alexander, Regional interpretation of water-quality845

monitoring data, Water Resources Research, 33 (12), 2781–2798, 1997.846

Smith, P. J., K. J. Beven, and J. A. Tawn, Detection of structural inadequacy in process-847

based hydrological models: A particle-filtering approach, Water Resources Research,848

44 (1), doi:10.1029/2006WR005205, 2008.849

Stonestrom, D. A., B. R. Scanlon, and L. Zhang, Introduction to special section on impacts850

of land use change on water resources, Water Resources Research, 45, 2009.851

U.S. Bureau of Census, U.S. census, available at http://www.census.gov, 2010.852

U.S. Geological Survey, Hydro 1k elevation derivative database, available at853

http://eros.usgs.gov/Find Data/Products and Data Available/gtopo30/hydro, 2000.854



U.S. Geological Survey, National Land Cover Database 2001 (NLCD 1992/2001 retrofit855

change product), U.S. Geological Survey, 2001.856



Reaches not included in set J(i)

Reaches J(i)

Upstream monitoring station

Reach i

Figure 1. Schematic illustration for SPARROW reaches where J(i) is the set of adjacent

reaches upstream of reach i.

R2 flux 0.885R2 yield 0.802Mean square error 0.404Number of observations 354

Table 1. Results of Model I (1992) calibration



< -0.75

-0.75 to 0

0 to 0.75

>0.75

LN Residuals

Under-predict Over-predict

1992

!(

!

(

(

((!(!((

!!!

!

!!

(!(

((((((!(( ((((

(

((!((

(((((

!

!(((!

!!!!(((!!(((

((

(!(

((

!!

!(

((!!!

!!!!

!!!!

( (((

!!( ( (

(((

(

!(!

(

!

(!!! (

((

!!((!!

((

(

(!(!!

!((((

(

!!!!!!

!

!

!

!(

((

!!(

( (

( ( (!

(! !(((!!

!!!(!!

!

!

!! ( !!!

!((

!!

(!(

!

((

!((

!!!((

(

!(!

!(!!! !

!

!

(!!

!!( !

!

!

!

((!

!

!((

(( !

!!

!

!((

(

!(!!((

!(

(

(

!(!

(!

(

!(!(

!!

!(!!(

!((!

!((!

(!!

!

(

!(!

!

!

!

!

(

!

!!

!

!

!!(!

!

(

!!

(!

!!!

!!!!!!!!

!(

!

(

(

(

(

(

!

!(

!(

!!

!!

!

!

!(

!

!

!

(

!

!!

(

!

(

(

!(

((

(

!

!

!

!

!

(

(

(

(

(

!!(!

!

!

(

!

(

!((

!

(

Over-predict Under-predict

8 10 12 14 16 18 20 22

Ln of Actual Flux

8

10

12

14

16

18

20

22

Ln

ofP

redi

cted

Flu

x R2 = 0.885

1992

Figure 2. Predicted flux from Model I (1992) versus actual flux of total nitrogen for

the 354 monitoring stations for 1992

< -0.75

-0.75 to 0

0 to 0.75

>0.75

LN Residuals

Under-predict Over-predict

2001

!(

(

!

(

(

!

!

!

( !

!( !

!!

!

!!

!(

!

!(

!

!

!

!

!

(

(

!

(

!

!

(((

!

!(!

(

!

!

(

!

!

(

!

(

(

!

!

(

(

(

!(

!(

!

!(

!

(

(

(

!(!(

(

!(

(!(

(

(

!(

!(

(

!(!(

!

!

!

!(

!

!( !

!

!(!(

(

!(

(!(

!(

!(

!(

!

(

!(

(

(

!

!(

(

!(

!(

(

(

!(!(

!(

(

(

(

!(

!(

!(

Over-predict Under-predict

8 10 12 14 16 18 20 22

Ln of Actual Flux

8

10

12

14

16

18

20

22

Ln

ofP

redi

cted

Flu

x R2 = 0.890

2001

Figure 3. Predicted flux in Model IV (2001) versus actual flux of total nitrogen for the

122 monitoring stations for 2001



Parameter Units Coefficient Standard Error P-Value

Nitrogen Sources, β

Population related Sources kg person−1 yr−1 3.565 0.801 <0.05Atmospheric deposition dimensionless 0.282 0.179 0.117

Fertilizer application dimensionless 0.248 0.046 <0.05

Livestock waste production dimensionless 0.083 0.074 0.260Non Agricultural Land kg km−2 yr−1 310.05 53.68 <0.05

Land to water delivery, α′

Soil permeability in hr−1 -0.097 0.017 <0.05Drainage density km−1 2.006 0.479 <0.05

Mean annual air temperature ◦C -0.061 0.009 <0.05

Precipitation cm 0.009 0.001 <0.05In-stream decay, κ′

κ1 (Q ≤ 28.3m3 s−1) day−1 0.226 0.031 <0.05

κ2 (28.3m3 s−1 < Q < 283m3 s−1) day−1 0.059 0.025 <0.05Reservoir decay, κr m yr−1 7.182 1.938 <0.05

Table 2. Resulting model coefficients for the SPARROW model from the Model I

(1992) calibration.

Incremental Nitrogen Yield 10th 25th 50th 75th 90th Mean Standard(kg ha−1 yr−1) Deviation

Model I (1992) 2.9 4.52 8.16 14.11 24.64 17.91 603.51

Model II 2.66 4.19 7.45 13.12 22.98 16.66 528.83

Model III (2001 Sources) 2.95 4.64 8.17 14.18 24.63 18.03 600.64

Model IV (2001) 2.71 4.3 7.49 13.2 22.91 16.8 525.7

Total Change (kg ha−1 yr−1) -0.19 -0.22 -0.67 -0.91 -1.73 -1.11 -

Model IV - Model I

Total Change (%) -6.55 -4.87 -8.21 -6.45 -7.02 -6.20 -

Model IV - Model I

Change due to Hydrology (%) -8.28 -7.30 -8.70 -7.02 -6.74 -6.98 -

Model II - Model I

Change due to Sources (%) 1.72 2.65 0.12 0.50 -0.04 0.67 -

Model III - Model I

Table 3. Distribution of incremental total nitrogen yield (kg ha−1 yr−1) for the different

model scenarios.



Change in Precipitation (%) Change in Temperature (°C)

Figure 4. Percent change in precipitation and temperature between the input datasets

used for the 1992 and 2001 models.



Incremental Yield Difference

2001-1992 (Percent)

a. 1992 Incremental TN Yield b. Inc. TN Yield difference due to Hydrology


Due to Hydrology (Percent)


Due to Sources (Percent)

c. Inc. TN Yield difference due to Sources d. Inc. TN Yield difference due to Hydrology

and Sources

Figure 5. The top left map (a.) shows incremental nitrogen yield scenario in 1992. The

top right map (b.) shows the percent difference in incremental nitrogen yield because of

change in hydrology between 1992 and 2001. The bottom left map (c.) shows the percent

difference in incremental nitrogen yield due to the change in source contribution between

1992 and 2001. The bottom right (d.) map shows the percent difference of incremental

nitrogen yield due to overall change between 1992 and 2001.



Population Atmospheric Fertilizer Livestock Waste Non-Agricultural0.0

0.5

1.0

1.5

2.0

2.5

3.0

Incr

emen

talT

NY

ield

(kg

ha−

1yr−

1)

Model I (1992)Model II (2001 Hydrology)Model III (2001 Sources)Model IV (2001)

Figure 6. Median incremental total nitrogen yield (kg ha−1 yr−1) for different sources

of nitrogen and for each model scenario.

Population Atmospheric Fertilizer

Non-Agricultural


2001 - 1992 (Percent)

Livestock Waste

Figure 7. Percent difference of incremental nitrogen yield between Model I (1992) and

Model IV (2001) for different sources of nitrogen



Population Atmospheric Fertilizer Livestock Non-Agr.0

5

10

15

20

25

30

35

40

45

Incr

em

enta

l Sourc

e S

hare

(%

)

Model I (1992)Model IV (2001)

Population Atmospheric Fertilizer Livestock Non-Agr.10

5

0

5

10

15

Perc

ent

Change in S

ourc

e S

hare

(1992 t

o 2

001)

Figure 8. Comparison of median incremental total nitrogen source share (%) contribu-

tion for Model I (1992) and Model IV (2001).



Distribution of Incremental Range of Yield values

Model No Watershed Type1 Yield (kg ha−1 yr−1) from Literature2

10th 25th 50th 75th 90th (kg ha−1 yr−1)

Model I Urban 12.61 17.82 32.2 66.65 113.75 1.6 - 38.5(1992) Forest 5.49 7.77 11.13 16.38 30.21 0.1 - 10.8

Crop 8.09 13.61 23.73 32 40.83 0.8 - 79.6

Grass 1.73 2.49 3.67 5.26 8.26 0.1 - 30.8

Model II Urban 12.16 15.17 29.81 60.31 119.84 1.6 - 38.5

(2001 Hydrology) Forest 5.26 7.42 10.43 16.09 30.72 0.1 - 10.8

Crop 7.67 12.75 20.81 27.29 34.91 0.8 - 79.6Grass 1.49 2.27 3.41 4.9 7.52 0.1 - 30.8

Model III Urban 13.46 19.24 34.38 55.35 105.06 1.6 - 38.5

(2001 Sources) Forest 5.62 7.82 11.08 16.37 29.92 0.1 - 10.8

Crop 7.91 14.21 23.23 30.93 41.63 0.8 - 79.6Grass 1.73 2.55 3.8 5.45 8.61 0.1 - 30.8

Model IV Urban 13 17.09 32.07 49.69 111.31 1.6 - 38.5

(2001) Forest 5 7.5 10.49 16 29.96 0.1 - 10.8Crop 7.72 13 19.89 26.21 35.31 0.8 - 79.6

Grass 2 2.32 3.53 5.1 7.86 0.1 - 30.8

Table 4. Incremental total nitrogen yield (kg ha−1 yr−1) for different model scenarios

and literature estimates for watersheds with dominate landuse type in the United States.

1The land cover types are based on the following percentage of land use area in SPARROW

watersheds: urban(>90%), forest (>95%), crop land (>90%), grass (>90%)

2 Literature reported values for incremental total nitrogen yield [Frink , 1991; Ritter , 1988;

Beaulac and Reckhow , 1982].



Urban Forest Crop land Grass0

5

10

15

20

25

30

35

Incr

emen

talT

NY

ield

(kg

ha−

1yr−

1)

Model I (1992)Model II (2001 Hydrology)Model III (2001 Sources)Model IV (2001)

Watershed Type 1

Figure 9. Median incremental total nitrogen yield (kg ha−1 yr−1) for major land use

types in the United States for different model scenario.

1The land cover types are based on the following percentage of land use area in SPARROW

watersheds: urban(>90%), forest (>95%), crop land (>90%), grass (>90%)


toward disentangling the e ect of hydrologic and...

Documents